Global took its cell counts from `global_cell_counts()` — one gridunit per
region — and discarded the requested extent entirely. On GJ380c that produced
a 191x95 canvas built from a heightmap stored at 512x256: roughly seven times
the available cells thrown away before anything was drawn. The count also
shrank as REGION_M grew, so tuning the scale ladder silently degraded the
opener, which is why the top of the ladder got worse rather than better as
the ladder itself was refined.
Global now fits the largest 2:1 canvas inside the requested extent. It cannot
take its ASPECT from the viewport — the canvas is equirectangular whole-body,
360 degrees of longitude by 180 of latitude, and must stay 2:1 or the cells
stop being square and the map shears — so the existing letterbox absorbs the
remainder. A hostile extent is still clamped; sizing to the request is not
trusting the request.
Global also joins the deep display ratio, making the band uniform. At 5 px
per gridunit a 1920 px window asked for 384 cells across a body whose
heightmap holds 512x256 — discarding stored detail to save work already done.
At 2 px it asks for 960, which is heightmap-native: nothing thrown away,
nothing invented, and the same screen area filled either way.
A body with no radius is not a sphere (asteroid belt, oort cloud) and has no
equirectangular surface to fit. Those degrade to the region grid — a visibly
degenerate 1x1 canvas — rather than a plausible-looking lie at whatever size
the viewport happened to ask for.
project.yaml 0.4.3 -> 0.4.4 invalidates the persisted step-canvas disk cache.
District/Quarter/Block/Chunk kept identical 960x540 cell counts through the
extent inversion, so their cache keys are byte-identical while a District
canvas now covers 3.6 km of ground instead of 1,966 km — a warm cache would
silently serve pre-inversion canvases.
Global still rides the orbital derive, so it carries no courses yet; that is
the next step and is deliberately separate, being a cost question over the
whole body rather than a sizing one.
Pair session with Jeroen, 2026-07-26.
Co-Authored-By: Claude <noreply@anthropic.com>
A rung used to fix the gridunit SPACING, with the canvas extent falling out
of spacing x cell count. That is why the top of the ladder was unusable: at
REGION_M spacing a viewport-sized canvas spanned ~251,658 km — six times
around a rocky body — so the Region rung capped to the body and redrew the
Global picture pixel-for-pixel. "Global and region look the same" was not a
rendering bug; it was this relation, stated in metres.
Inverted: a rung fixes the EXTENT and the spacing falls out of the canvas
size. The shorter viewport axis spans exactly one cell of the rung's level,
so a widescreen window shows more ground on the long axis rather than less
on the short one. Every rung now shows the ground its name promises —
Region 262x466 km, District 4.1x7.3 km — and the canvas cell count is
viewport-driven and identical at every rung, so derive cost no longer varies
with depth and resize is free.
Consequences that fell out of the inversion rather than being chosen:
- Region leaves the orbital derive set. It was envelope-only because at
251,658 km nothing finer made sense; at 262 km it is a genuine provincial
map and takes the full courses-aware derive. Region having no rivers at
all was much of why the top of the ladder read flat. It also joins the
deep display ratio for the same reason.
- The S2 station-spacing floor is deleted, not retuned. It guarded an
O(1/spacing) blowup that the inversion makes structurally impossible (the
canvas cell count is now constant across rungs, so stations-per-course is
bounded however deep you scroll). Kept, it would do active harm in the
opposite direction: a 2,048 m pitch across a 3.6 km District canvas places
two stations and draws every river as a straight line. Station placement
gets its own generator pass.
- cap_extent_to_body is superseded and now a documented no-op. A canvas can
no longer over-request a body by construction. The residual question —
whether a rung's cell exceeds the whole body — is liveness, not capping,
and is_rung_live_on_body() answers it by omitting the rung. Empirically it
never fires on inhabited content: all six rungs are live on all 271
populated bodies with a radius.
- snap_to_gridunit no longer truncates its multiplier to int. Post-inversion
the deep rungs run sub-metre (Chunk ~0.12 m at a 1080 px short axis), where
int(spacing) floors to zero and would collapse every request centre onto
the origin.
Both sides derive spacing from the same three inputs (rung, echoed cell
extent, body radius) rather than one telling the other, so there is nothing
to keep in sync beyond the constant table itself. Body radius already
reaches the viewer via enter(); no wire change.
Pair session with Jeroen, 2026-07-26. D-243/D-255 amendments to be backfiled.
Co-Authored-By: Claude <noreply@anthropic.com>
Two viewer fixes from the same pair session; they share
step_canvas_viewer.gd so they land together.
FETCH READOUT. A cold derive takes seconds and the map gave no honest
sign of it. Root cause found while building the replacement: a Control
paints its own _draw() BEFORE its children, so everything the viewer
drew itself — the old DERIVING TERRAIN label AND the pending wash —
was painted UNDER the terrain canvas, visible only when no texture
existed at all, i.e. never in the slow-fetch case they existed for.
That has been the state since the stepped viewer shipped. The readout
now lives in its own overlay node added after the canvas: a centered
implant-idiom panel, DOWNLOADING MAP DATA, indeterminate sweep, 250 ms
grace so cache hits never flash it, held canvas still drawing beneath.
Indeterminate by design — the server reports no derive sub-steps, and
a progress fraction we cannot source would be invented.
ONE SETTLE TIMER, THREE TRIGGERS. The viewer never re-requested a
canvas on resize, so one derived for a smaller window letterboxed
forever in a bigger one — the map not filling the frame. And the
pan-edge refetch fired from inside the per-frame pan loop the instant
its threshold was crossed, so a held edge-scroll issued a fresh
request AND snapped the view on every frame past it. Both are the same
event: the user is still moving. A shared one-shot timer now collapses
them into a single request at the resting state, with a hard pan
threshold that still fires immediately when the canvas edge is about
to enter view (waiting there would show empty background), and
drifting back inside the soft threshold cancels the pending request.
Resize reaches this Control identically whether the OS window or a
diegetic in-implant parent changed, so both sources are covered.
Two new tests pin the soft path (schedules, does not refloat or snap)
and its cancellation; the pre-existing threshold test was verified to
still discriminate — its drift trips the new hard threshold — rather
than passing vacuously.
Co-Authored-By: Claude <noreply@anthropic.com>
DISPLAY_RATIO_DEEP 1.0 -> 2.0, so a viewport-fit request at District/
Quarter/Block/Chunk asks for ~4x fewer gridunits (1920x1080: ~2.07M
cells -> ~518K) and the server-side derive cost falls with it. Texel-
exactness is preserved — the ratio stays a whole number of screen px
per gridunit, so every source texel still lands on whole pixels; only
the block size changes. Non-integer ratios are not an option here:
they reintroduce exactly the sub-pixel blur D-255's texel-exactness
exists to prevent.
Judged live by Jeroen against the 1x1 build (pair session): the deep
rungs read as crisp larger pixels rather than blur, and the speedup is
substantial. Looks were the gate.
Both transport tests that pinned the old literal now assert the
RELATIONSHIP instead — deep rungs share the constant, viewport-fit
divides by it — plus a new guard that the ratio stays whole, so the
value remains tunable without editing tests that are not about it.
Co-Authored-By: Claude <noreply@anthropic.com>
The legend's reposition() now derives its Y from the header panel's
MEASURED bottom plus a named 12px gap (new accessors on the viewer;
deferred recompute so it reads settled layout, refreshed on header
content change) — the old hardcoded 60.0 sat 17px inside the wrapped
header's real 77px bottom, fusing the panels. Pixel-proven at the
evidence center: bottom=77.0, legend top=89.0, gap=12.0 exact. All 13
goldens regenerated at the DOCUMENTED 960x540 (the 1280x720 rider was
my own CLI flag, not a config; PIL-verified) with two-run byte
stability. Two structural tests pin the header/legend relationship
(T-1192 precedent), proven failing-first against the old constant
(-17.0px reported), and the legacy test that pinned the literal 60.0
now asserts the derived relationship so it cannot re-enforce drift.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Root cause was not the suspected fade race — no animation exists
anywhere under client/ui/implant (grep-proven, and settle+2 vs
settle+90 captures were byte-identical). _build_screen_header() was
the single ImplantHeader call site in the codebase that added the
header bare instead of through ImplantPanel.add_component(), so its
fixed light-gray text washed out against pale terrain (Quarter upland
scatter, District olive) while reading fine on dark ocean. Wrapped in
an ImplantPanel like every sibling — the theme's panel_bg scrim makes
it legible everywhere. One committed golden (atlas_GJ820Bc_District)
already carried the ghosting baked in, confirming the bug was static;
all 13 goldens regenerated per the header change, two independent
live-server runs byte-identical (T-1157 stability discipline).
Verified live at the original evidence center: header crisp at
settle+2/+30/+90.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
user://atlas_cache/ is one directory for every worktree gate run, live
capture driver, and real play session, and T-1183's Tier-2/3 lookup
short-circuits BEFORE test_mode's silent-no-op IPC — so a warm shared
cache delivers real canvases into tests written against 'nothing ever
arrives'. Caught live: a concurrent GJ380c Global capture flipped the
two before-any-canvas viewer tests in another worktree's push gate.
Adds disk_cache_root_override on StepCanvasViewer (threads into
StepCanvasRequest's existing test-injection-only seam; production
leaves it empty) and routes every direct viewer construction in the
viewer + legend suites through a _make_viewer() helper pinning an
isolated root. First slice of T-1193's viewer test seam; screen-built
viewers still share the default root (noted on T-1193).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Every screen-targeted intent now routes through one
_require_current_screen() check and returns the structured error shape
instead of silently mutating an off-screen viewer (hoshe's finding:
scroll_rung from the reach screen fired real IPC and reported ok). The
reference driver's fixed 4-frame settle becomes is_pending()-aware with
a 600-frame bound, the keep-waiting decision extracted as a pure
testable function — restoring the proven eyeball-driver discipline. The
terrain_reference guard moves into AtlasApp._on_body_selected(), the
shared tail for double-click, Enter, AND the intent path — closing a
pre-existing click/Enter divergence hoshe caught this PR formalizing;
the intent layer pre-checks via the new SystemScreen.find_body() and
reports structured errors for unknown ids and terrain-less bodies.
after_test() resets AtlasAgentBridge.current_app (tyre's freed-pending
footgun). Suites 58/58 + 14/14; full suite 3,638.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
D-226 layer 4, rebuilt against the post-D-255 stepped Atlas after the
Phase-1 reconciliation (the original intent list targeted the retired
continuous-zoom viewer). Eleven intents, each backed by the exact
production handler a click calls — select/open for systems and bodies
(extracted shared by-id tails so click and intent paths are one code
path), scroll_rung, reset_view, back, open/close_atlas, set_overlay —
plus two new first-class capabilities: jump_to_center (the fixed-center
revisit pattern proven by five eyeball drivers, via a new
StepCanvasViewer.jump_to seam that reuses _scroll_rung's exact request
tail — same extent cap, same cache keys) and get_current_canvas_summary
(allocation-light reads off the raw wire dict, courses-by-class,
draw-matched settlement dedup — no PNG decode). Contract shape: dumb
AtlasAgentBridge autoload holding the app handle (untyped per the
parse-order rule), all logic in the static AtlasAgentInterface class.
observe() is side-effect-free: current state + a generic Control-walk
affordance tree. In-process consumers only this ticket (documented);
the committed reference driver (atlas_agent_driver.gd, InputSwallower +
settle-until-ready from the T-1157 inventory) replaces the scratch
eyeball drivers as the sanctioned headless-drive pattern.
select_city/open_regional dropped with recorded rationale (no
settlement hit-test affordance exists post-D-255) — diff on the ticket.
33 new tests across three suites.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
One commit for two tickets whose changes share the bridge/plugin
plumbing files. T-1169 connects the three dormant feature-name pieces:
atlas_feature_names populated at regen (17,891 rows — 15,190 mountain,
2,701 river — via populate_atlas_feature_names mirroring the city-names
importer; systems.db regenerated, stamp fresh), attach_feature_names
wired into the cascade's Topography block with name pools threaded
DB-free through AnalyzeBody (D-225 pattern) and assignments stored on
Layer1Output/BodyWorldState for future consumers, and a
FeatureNamesRequest/Response read proxy as the bridge's 7th tagged
envelope (D-236 pattern, both SimBridge impls). Client label DRAW is
deliberately NOT here — implementation proved both river and mountain
labels need a wire-carried position (the pool is position-free; course
polylines aren't correlated with the named attractors by construction) —
deferred to T-1195's single design pass. cascade_layer1 golden re-pinned
(additive feature_names field).
T-1159 retires the legacy u32 granularity field fully shadowed by
window_granularity_v2: AtlasLayerRequest.window_granularity,
DistrictWindowLayer.granularity echo, the u32::MAX sentinel, and
resolve_window_granularity are gone server-side; client encode paths and
the caller-less atlas_window_cache legacy key component dropped;
msgpack fixtures regenerated; the T-1150 aliasing regression test now
drives through the surviving enum field. The district_window carrier
itself survives byte-compatible per D-255(c).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Five findings from the araminta+hoshe round. Tapering now fires only on
a TRUE upstream source: the course's first raw world point is tested
against the canvas's own world bounds (conservative 1m epsilon) —
exact detection because the server crop keeps one point beyond the
window (layer_proxy crop_course_to_window lo = first_in-1, contract
documented), so crop passthroughs draw the old flat full-width cut and
never a false headwater. The averaged-normal joint is replaced by a
real mitre (half_w/cos(theta/2) recovered trig-free via the bisector
normal), clamped by a 2x mitre limit AND 0.45x the shorter adjacent
segment — restoring true perpendicular width at bends (the 29% pinch at
confluences is gone) and preventing the hairpin bowtie; the winding doc
now states the actual bounded guarantee. Antialiasing restored via the
hybrid: only the varying-width taper head draws as a ribbon; the
constant-width ~85% of every course keeps the original antialiased
draw_polyline (byte-identical for untapered courses), split at an
interpolated arc-length point sharing position and width — junction
capture evidence in .cache/screenshots/t1175-fix-round/. Flat-fill
single-element color array. Suite 24 -> 48 tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The single large polish pass Jeroen requested off the T-1170 captures,
benchmarked against RimWorld's world-map fluency. Courses now draw as
source-tapered ribbons: per-vertex width ramps from a hairline at the
upstream source to full class width over 15% of the course's arc length
(arc-length parameterized, not vertex-indexed, so point density doesn't
change the read), built as one draw_polygon ribbon with mitred joins.
Class width table retuned 0.9/1.4/2.2 -> 0.6/1.2/2.4: a clean ~2x
per-class ladder so a tributary-joins-trunk confluence reads as a join,
trunk held at its visually-proven weight, stream thinner per the
benchmark's thin/consistent/restrained grammar. Opacity untouched
(Araminta's T-1170 ruling stands). Coast-gradient item resolved as
already-correct (the coastal transition zones + elevation lightness
render the shoreline band; capture-verified) — no wire change. Stipple
assessment filed as T-1194. Eight new taper-geometry tests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The fractional fit branch is deleted, resolving both review findings at
the root: tyre showed its comments cited D-255 for an exception the
record does not contain (the language came from the lead's ticket text,
not governance), and hoshe showed it returned sub-1x for a canvas
exceeding the viewport on one axis. Replacement: fit_scale_ratio()
chooses the largest integer pixels-per-gridunit R fitting both
legend-reserved axes, floored at 1 (over-viewport draws native and
crops like every fixed rung) — the fine-grained integer lattice (GJ1c
1080p -> 9px/gu = 1593x792, ~98% width; 4K -> 20) that makes the
fractional hatch unnecessary. center_offset() floors to whole pixels
(half-pixel centering would blur the texel grid). Doc comments cite the
real sanction (D-255 amendment 2026-07-25, this branch). Legend column
constant is now canonical in transport, read directly by the legend
(was an independently-typed literal); its test asserts real geometry.
New regression test proves _global_body_extent clears on body switch
and never caps another body's requests.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Lake edges rendered as hard step-edges while ocean coasts got multi-tone
transition bands: every coastal-transition morphology gate keys on
ocean_fraction_q, definitionally 0 inside a lake basin (hypothesis (b)
of the ticket; (a) disproven first — a shoreline-crossing sweep at
2048/512/128m plus a 10m fine sweep all land on the same continuous
crossing, so positional refinement was never broken). New
DistrictProfile.lake_margin_q (0-100 settled-hydrology depth band, from
the same bilinear filled/elevation pair the lake test already samples;
ceiling calibrated just above the observed p90 depth on GJ338Bd's 5,043
flooded cells), threaded through both derive paths onto
EncodedStepCanvas (serde-default for shape tolerance) and down the
client: protocol decode, terrain-layer plane, colorize shades Lake cells
by depth band instead of elev_q (bedrock-under-water, the wrong signal).
project.yaml 0.4.0 -> 0.4.1: the new wire field must invalidate the
client disk cache via its version tag (T-1183's D-192 mechanism).
Acceptance gates green with the new field (lossless round-trip,
cache-hit==cache-miss, every rung).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The two client fixes for the Atlas frames Jeroen flagged. Region rung no
longer requests more planet than exists: cap_extent_to_body() caps the
viewport-fit extent at the body's own region grid (the Global canvas
extent echo — cols=regions_per_equator, rows=cols/2), matched by gridunit
SPACING not rung name, applied before the request/cache key is built;
cold-start scroll-before-Global-echo requests uncapped and lets the
server clamp (documented). Kills both the side-by-side continent repeat
and the past-the-pole stripe smear.
Global opener now fit-scales and centers through one shared letterbox
mechanism (center_offset/integer_fit_scale/fit_scale, 75%-coverage
integer-vs-fractional decision, NEAREST already forced on orbital rungs
per D-255's escape hatch) also used for the Region cap's letterbox
remainder. Legend column reserved before fitting — pinned to the legend's
own width by a cross-constant test, never overlapping the canvas.
Also fixes a latent crash: NOTIFICATION_RESIZED fires mid-_ready() before
children exist; null guard in the resize path.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Atomic index writes (tmp+rename), orphan-payload reconciliation sweep
folded into the background sweep, payload shape guard mirroring the
protocol's own width/height discriminator, size_bytes via get_position()
instead of a full payload re-read, and a 64-bit SHA-256 payload filename
(String.hash()'s 31-bit space made a silent wrong-map filename collision
a ~1-in-16k event per cap-full body; migration self-heals via the orphan
sweep). Tier-3 class doc rewritten to name the real D-253 seam — the
glaciation/flooded_q wire fields exist but carry static values and no
staleness signal crosses the wire; T-1190 tracks threading a wire TTL
into put(sim_ttl_sec) when the driving clock lands — replacing the false
'no sim-state wire field exists' premise. Four new tests plus a
_stub_canvas fixture helper centralizing the canvas shape contract.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
step_canvas_disk_cache.gd: the Tier 2/3 store per D-255(d) and the
round-2 three-tier spec. Payload is the WIRE, pre-decode — store_var/
get_var round-trips the PNG-encoded PackedByteArrays natively, and
Image.load_png_from_buffer never runs in this file. One composite key
shared verbatim with Tier 1 (hash filenames for filesystem safety);
per-body index.json with malformed-index recovery (rebuild-or-discard,
never crash).
Three independent eviction mechanisms, exactly as ruled: rung-0 Global
carries a retention floor no sweep touches (now also threaded into
Tier 1 per the T-1182 handoff); Tier 2 geometry is byte-valid forever
and evicts only by time-since-last-visit (14d starting tunable, on
body-open) and LRU byte budget (256 MiB/body, 5-min coarse timer) —
two separate sweeps; Tier 3 sim-state TTL is wired and tested but has
no production caller yet (no sim-state field exists on
EncodedStepCanvas — the D-253 stub inheritance, documented).
Hardening per D-255(d), both mandatory: per-body deep-rung cap
(512 Block+Chunk entries, enforced synchronously in put(), floor- and
budget-independent — the ticket sanctions count-or-quota; count chosen
as the direct D-226(d) information-content proxy) and a schema/version
tag on every entry (project.yaml version via the existing
loading_screen line-scan idiom — D-192 co-ship makes the client
version the wire-schema version; exact-inequality mismatch = miss +
drop, NEVER decode, checked in both has() and get_canvas()).
Integration: request_now() checks Tier 2 on a Tier-1 miss (synchronous
promote), Ready responses write through to both tiers, Pending never
writes; viewer runs the visit sweep on body-open + the background
sweep on a 5-min timer.
30 new disk-cache tests + 7 request-integration + 3 sweep-wiring tests
(restart persistence, sweep independence both directions, cap
semantics, version-mismatch never-decode, corrupt-index recovery).
Full client suite 3,440/3,440, 0 orphans; cold-parse clean.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Tyre finding: the orphaned AtlasViewer cluster is now actually deleted
(atlas_viewer, atlas_marker_overlay, atlas_descend_geometry,
atlas_legend_panel — a sixth orphan found beyond the review list —
atlas_generation_proxy, atlas_generation_state; ~2,497 lines), with
reachability re-verified across preload/class_name/res:// strings,
every .tscn, and the standalone companion app. Test suites triaged,
not blanket-deleted: 5 pure AtlasOverlayColors tests relocated into
test_atlas_window_colors, the live type-identity regression guard
relocated into test_step_canvas_viewer, dead coverage deleted. A real
harness gap surfaced during diligence and RULED, not patched:
visual_scenarios/visual_capture golden shots call retired
continuous-zoom API — no shim (would resurrect what D-255 kills);
inventory recorded on re-scoped T-1157 (gate-invisible, manual
targets only).
Hoshe finding 1: decode_png_field now detects the [Error,
PackedByteArray] bin-shape from messagepack.gd explicitly — a genuine
msgpack bin payload decodes correctly instead of silently collapsing
to [0,0]; test built from a real round-tripped bin decode.
Hoshe finding 2: the hard zoom-out reset is wired — ascend at rung 0
with a drifted view triggers _reset_to_global (the restored HARD
condition), behavioral tests through the real input path.
Notes folded: refloat + edge-scroll test coverage, legend smoke suite,
Vector2i narrowing-safety comment with computed headroom.
gdlint clean on touched files; full client suite 3364/3364.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The two-layer client rebuild per D-255(a)(b)(e), replacing the
_canvas.scale continuous-zoom model with one viewer, one path, all six
rungs:
- step_canvas_protocol.gd: StepCanvasRequest/Response codec against
the T-1181 wire contract — incl. the discovered png_bytes subtlety
(rmp_serde without serde_bytes emits a msgpack int-array, not bin;
decode repacks via PackedByteArray before load_png_from_buffer) and
the extent-echo rule (read the server-clamped extent, never assume
the requested one).
- step_canvas/ component: transport (six-rung ladder, cursor-anchored
scroll steps, edge-scroll/WASD pan with re-request on edge crossing,
hard reset-to-Global), RTT terrain layer (Image.set_pixel colorize
per the c1 measured ruling, texture.update reuse on step-cross,
NEAREST coarse / LINEAR fine per rung), unscaled screen-space
annotation sibling (courses + settlement markers at literal px),
in-memory LRU cache (Tier 1; T-1183 layers the disk tiers beneath),
request lifecycle (pending retry, staleness gate, extent echo).
- Full _canvas.scale retirement in the same change: the zoom-scaled
canvas model, the _zs compensation family, select_rung /
MAX_COVERAGE_M / compute_tile_grid, the orbital-mosaic-vs-window
two-path split, _view_zoom/_canonical_fit_zoom — 10 source files
deleted; their 14 test suites deleted with them (T-1157 dead-goldens
rule; replacement visual-capture coverage is re-scoped T-1157).
- Surviving surfaces kept per the ticket: atlas_window_cache.gd's LRU
shape (the ticket's named file atlas_window_tile_set.gd was the
retiring orchestrator; the real LRU shape lives in
atlas_window_cache.gd — cited in step_canvas_cache.gd), overlay
colors, legend/overlay-bar chrome, AtlasViewer descend geometry.
Determinism boundary per D-255(e): the client interpolates only within
the closed server-supplied input set. 7 new gdUnit suites (164 cases)
incl. a real extent-echo bug caught by its own test during
implementation. Full client suite green (exit 0) with the live-gated
suites running against a worktree server build.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Two separable findings from the live A/B (real driver): (1) the
original zero-ring report was a viewport-framing crop — the ring
fired all along at screen (1755,-191), above the frame under the
COVER fit; recentering via set_view proved the path live. (2) The
real bug once in-frame: at Quarter fit zoom the compensated ring
radius (5.0/7.5=0.667 canvas) fell below the floored stroke (1.0
canvas) and draw_arc's stroke filled its own hole — a solid blob,
not a ring. Bracket: 1x stroke=blob, 1.5x=hollow recovers, 2x=clean.
Fix: zoom_compensated_ring_radius() floors the radius at 2x the
paired stroke, wired through _zs_ring_radius() for both ring and
halo arcs; verified live producing a clean hollow double-ring. The
third member of the Godot sub-canvas-unit rasterizer family (width
floor, stroke-vs-width sites, now radius-vs-stroke) — all recorded
for the T-1176 render-mechanism discussion. 10 pin tests in the
stroke-width suite; revert-verified (floor drop -> named failures);
full client suite 3956/3956; gdlint clean.
Tickets: T-1170
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Her audit of the course captures found a uniform 1px alpha-255 line
where trunk should draw 2.2 screen px. Live A/B bracket (20/1.5/1.1/
0.6 forced widths + draw_line-vs-draw_polyline control) localized the
cause: Godot's line rasterizer floors stroke widths below ~1.0 canvas
units to hairline — the _zs arithmetic was correct (1.4/3.75=0.373
round-trips exactly); the value died at the driver. Fix: zoom_
compensated_stroke_width() (the _zs divide maxf'd at 1.0) via a
_zs_stroke() wrapper on EVERY stroke-width site (course polyline,
skeleton chords, mouth-ring arcs, basin boundary, attractor outlines
— same latent class everywhere even where not yet visible); radius
args proven unaffected and left on _zs. Honest degradation direction
documented: at high zoom effective width grows rather than pinning.
Class question resolved with printed ground truth: the original
window is genuinely single-drawable-class (trunk course degenerate at
1 point); a confluence window confirms real multi-class rendering.
The drive now prints per-course class/width/effective-px tables every
run. New 14-test stroke-width suite (own file, line-cap split) pins
the exact floor-engagement numbers and the PR#195-shape effective-
width >= table-value invariant; revert-verified (floor drop -> 4
named failures). Full client suite 3942/3942; gdlint clean.
Tickets: T-1170
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
_draw() splits into two independent gates: the Layer-1-gated skeleton
path (Region chords, clip retained) and the NEW DistrictWindowLayer-
gated course path — build_course_render_plan() (pure, render-free-
testable) consumed by draw_polyline with _zs-compensated widths and
opacities from the Araminta revisit tables; mouth double-rings at
Mouth termini only (EdgeDrain/ContinuesBeyondWindow/None: three
meanings, one presentation — draw to last point, stop, documented);
zero water clip on the course path by construction (courses carry
rung-consistent termini). CourseTerminus wire vocabulary kept re-
pointable pending A2's real serde names; synthetic Ruling-3h fixtures
mean the suites need zero changes when the server payload lands. Real
gap found and fixed: window arrival never redrew the nature overlay
after the first fit (one line in _on_window_ready — courses would
miss every window swap post-pan). Water-clip header rewritten to
RESTRUCTURED status (retired on course rungs; permanent at Region
until Region goes windowed, T-1143 ruling 2). Revert-verified
(visibility-gate bypass -> 4 named failures). geometry-nature
112/112, nature-overlay 58/58, viewer 78/78, zero collateral; full
sweep 3928/3928 after the full-import bootstrap; gdlint clean (viewer
1016->1017, pre-existing overage rides T-1158).
Tickets: T-1170
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
B1 (Ruling 5c): nature block split out of atlas_window_geometry.gd
(954/1000 cap pressure) into atlas_window_geometry_nature.gd; RIVER_
CLASS_VISIBLE_BY_RUNG replaced by SKELETON_CLASS_VISIBLE_BY_RUNG
(Region-only now) + COURSE_CLASS_VISIBLE_BY_RUNG (District trunk+
tributary; Quarter all three — the pre-announced Quarter-rivers-
return) with width/opacity companion tables as Araminta's single
revisit point; deliberately opposite unknown-tag fallbacks per reader
(skeleton->full, course->empty), documented.
B2 (Ruling 5a): Region dot-scatter upgraded to connected chords via
river_downstream — D8 direction decode (0-7 into drainage.rs's
(row,col) delta table, antimeridian wrap-aware), sentinel chain ends
(MOUTH=8 ring-on-land, EDGE_DRAIN=9 no ring, TERMINAL=10 reserved,
decodes like EDGE_DRAIN so the future endorheic server needs no
client change). Pure build_skeleton_chords() split from drawing for
render-free testability. Chord clip rule (3g pick): segment clips if
either endpoint OR midpoint is drawn water — three-point catches both
narrow-inlet and long-chord failure modes at one extra lookup;
documented. Self-caught during build: first draft misdecoded the
pointer as a river_cells INDEX; rewired to direction decode against
A1's real convention before leaving the branch. Dual revert-verified
(direction sabotage -> 6 named failures incl. the chain-threading
pin; midpoint-drop -> exactly the 1 named clip test). Suites:
geometry-nature 86/86, geometry 130/130, nature-overlay 24/24; full
sweep 3892 with only the 6 known pre-existing garment/gait failures
untouched by this batch; gdlint clean.
Tickets: T-1170
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Live round: the first clip removed only 133 pixels. Root cause (dossier-
traced): resolve_morphology_zone wrapped the QUERY toward a tile's raw
canonical center — the inverse of the painter's draw_col =
nearest_wrap_image(tile_center, held_center) — so seam tiles tested
containment against the wrong wrap-image and read real-but-wrong-
location land cells. Fix mirrors the painter exactly: wrap the tile's
own center toward held_center, test the (already held-wrapped) query
against that. Index math itself was confirmed correct end-to-end and
is now factored into shared AtlasWindowGeometry.cell_index_for_local_
offset() (full painter unification not applicable — the painter only
iterates forward, never reverse-looks-up; documented). Post-fix: 1082
traced positions cross-checked against actual painted pixels, 0 real
mismatches; regenerated captures show every dot/mouth on land. The one
pre-existing test encoding the buggy direction as correct was replaced
by positive+negative wrap-semantics tests (the negative one is the
reliable revert discriminator) plus a wire-accurate seam-tile fixture
with an honestly-documented proof limit (a single-tile fixture cannot
distinguish the wrap directions; the multi-tile scan tests can).
Also: two harness traps found and fixed in the lead's drive scratch —
canvas_items stretch factor (root.size now matches the 1920x1080 base
viewport) and mid-arrival snapshots (explicit is_fully_arrived wait;
the round-1 faint result was partly a 4-of-6-tiles capture). 46/46 +
50/50 + five sibling suites regression-free; gdlint clean.
Tickets: T-1172
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Jeroen's hands-on report: rivers continuing under the ocean. Tyre's
ruling implemented: the skeleton is rung-independent, the drawn coast
is rung-indexed (warp cutoff admits more octaves per rung), so
reconciliation is a presentation-frame operation — a draw-time clip
against the SAME per-cell morphology verdict the terrain painter used,
at the rung on screen. New pure module atlas_window_water_clip.gd:
cell resolution across both paths (single-window direct; tile mode
selects the containing tile by each tile's OWN echoed n with nearest-
wrap re-expression against that tile's canonical center — response-is-
source-of-truth + wrap discipline reused, not reinvented). Strict drop
(no snap); offshore mouths suppressed (return with real termini in
T-1170 — retirement markers at the clip sites); basins untouched;
fail-open wherever no composite data has arrived (the clip refines
presentation, never gates data). The _pos split feeds the SAME wrap-
resolved district to both draw position and clip test so they can
never disagree about the wrap image. 49 new tests incl. antimeridian
and mid-progressive-arrival fail-open; revert-verified with precise
attribution (breaking water detection fails exactly the 3 water tests,
fail-open/land tests stay green); smoke-suite stub crash under a real
driver caught and fixed (headless skip-gating masked it). 7 suites
regression-free; gdlint clean.
Tickets: T-1172
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
I1: _draw_attractor_shape's three stroke-width args (Confluence arc,
Coastal/NaturalHarbor arc, Oasis spokes) were raw screen-space literals
— Godot multiplies stroke widths by canvas scale exactly like radii, so
at the Region orbital fit zoom the outlines rasterized at ~0.01px, the
identical sub-pixel class the dot/ring compensation fixed, missed on
glyph internals (and attractors are Region-only — precisely where it
bites). Widths now arrive pre-compensated via a px_w param, keeping the
primitive pure. Regression pin: a source-scan test asserting no
draw_arc/draw_line in the function carries a bare numeric width (the
draw-smoke suite documents its own vacuous-pass mode, so source-scan is
the environment-independent gate); revert-verified by name. I2: D-226
visibility-direction sentence — Araminta's fade-down inversion recorded
as pre-T-1170 with its single revisit point named. I3: class-header
call-site claim corrected (enter() funnels through _enter_at_rung).
Tickets: T-1156
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
At 1.6px/60% the trunk dot under the mouth ring read as 'mouth glyph
over plain terrain' — invisible without knowing where to look,
underselling the ruling's own 'a major river crosses near here'
intent. 2.0px/80% preserves the fade-down ladder vs Region's
2.2px/100% without reading as accidentally-erased. Doc comment
records the revision provenance.
Tickets: T-1156
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Live round 2 (the real bug): every nature-overlay marker size was a raw
screen-space constant drawn inside _canvas, whose scale IS view_zoom —
at Lendel's orbital fit zoom (0.0063) a 2.2px trunk dot rendered at
~0.014px, invisible; the same code at District's 3.75 zoom produced the
correctly-visible mouth ring, which is why one capture worked and the
headline rung didn't. Fixed via AtlasWindowGeometry.zoom_compensated_
size() (pure, floor-guarded) wired through every radius/line-width;
basin FILL points are positions and correctly stay unscaled. Suspect
tile-mode-rung-detection was ruled out live (granularity_v2=Region
confirmed in tile mode) but pinned with a named regression test anyway.
+8 pure-function tests incl. a numeric pin of the pre-fix magnitude
(<0.02px at orbital zoom); revert-verified by name. Draw-smoke suite
documented as supplementary (the shared SubViewport background harness
can pass vacuously under X11 BadMatch — the pure suite is the gate).
Live round 3 (drive-script bug, no product change): the lead's scratch
drive passed the button LABEL to set_overlay_visible() and the unknown-
id guard silently no-op'd — but the chase banked a real pin:
test_set_overlay_visible_gen_basins_flips_gate_and_redraws_nature_
overlay (draw-counting spy per the cold-start precedent; is_queued_for_
redraw does not exist in this build). Revert-verified.
Basins verified live: 7 Lendel watershed boundaries render at the
ruling's alphas. Suites: viewer 76/76, geometry-nature 42/42, nature-
overlay 22/22, zoom-ladder 50/50, no regressions across the cluster.
Tickets: T-1156
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
New AtlasWindowNatureOverlay (Node2D on the viewer canvas, above
terrain): self-connects to the shared atlas_layers_received broadcast
(the established one-signal-N-consumers shape) and consumes the whole-
body layer1 skeleton per Tyre's carrier ruling — no windowed wire
touched. Coordinate chain layer1_pixel_to_world_m -> world_m_to_
district -> canvas-local lives in atlas_window_geometry as pure tested
functions; vertical convention (row 0 = North pole, wy increases
south) verified against the server's own pixel_to_world_m, pinned by
pole tests and revert-verified (sign flip fails them by name).
Araminta's per-rung presentation table implemented exactly: Region
full skeleton (radii 0.9/1.4/2.2, confluence 3.5, mouth double-ring),
District trunk-only 1.6px at 60% with mouths at full landmark styling,
Quarter off until T-1170 course invention; retired palette reused
verbatim; RVR on / BAS off / ATR off defaults; missing river_class
falls back to trunk. 58 new tests (34 geometry, 22 lifecycle, 2 real-
driver draw smoke actually rendered); 11 existing atlas suites
regression-free. atlas_window_viewer.gd runs 16 lines past the
advisory 1000-line cap on trivial wiring — accepted, rides T-1158's
decomposition.
Tickets: T-1156
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Araminta's ruling (PR #192 follow-up): filter keyed on rung IDENTITY via
the window's own echoed granularity_v2 — Region (incl. the orbital tile
mosaic, whose tiles are all Region-rung requests) samples NEAREST because
GPU bilinear at 204.8 km/cell reads as smoothing-over-absence; District/
Quarter keep LINEAR where cell density earns the blend. One shared helper
(_filter_for_granularity_v2) at both draw call sites; unknown/missing
wire tags fall back to LINEAR (never trusted into NEAREST). COMPOSITE_
SMOOTH survives as the independent compile-time pipeline axis — the
two-axes split is documented in the file header. Washes/border fades
untouched per the ruling. Focused suite 44/44; revert-verified (helper
hardcoded LINEAR -> exactly the three Region-NEAREST tests fail).
Tickets: T-1161
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Tyre's APPROVE items (T-1163): record the two-legal-wire-shapes-for-one-
logical-state protocol invariant (whole-response Pending AND Ready+null
district_window both mean 're-poll'; only NotFound/Error are terminal) as
a D-226 note under the T-1124 §4 amendment area, so a future server
refactor of the asymmetry must migrate every consumer in the same change.
Drop the duplicated 5-line stagger comment in atlas_window_tile_set.gd.
Follow-up tickets filed on main: T-1164 (tiled terminal-recovery),
T-1165 (queue_redraw edge root-cause), T-1166 (cold-launch test tier).
Tickets: T-1163
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The true cause of the black first launch, server-confirmed after four
disproven theories: a stone-cold body answers the FIRST window request
with a whole-response status 'Pending' (only the whole-body cache-hit
branch sets Ready), and on_response()'s very first check — status !=
Ready -> return — swallowed it before the retry machinery could run.
retries stayed 0 forever; the DERIVING state never resolved. Second
connections worked by luck (the first request warms the whole-body
cache, so they read Ready and take the healthy path). Both tile fan-out
AND single-window first-descents were affected — one shared function,
one fix: branch on the outer status first (the atlas_generation_proxy
reference shape): Pending -> retry, Ready -> existing null-window retry,
NotFound/Error -> give up immediately (the principled give-up policy,
replacing the elapsed-retries ceiling).
Hardening in the same round: deterministic exponential backoff (0.5s
doubling, 4s cap) + per-tile stagger (0.1s * index — six tiles retry at
0.5/0.6/0.7/0.8/0.9/1.0s, strictly-increasing asserted, not jittered);
MAX_RETRIES 20->30 (~110s horizon under backoff).
Regressions are wire-accurate by construction: whole-response Pending
(body_id only, no center/n/granularity — verified against the server's
own response construction) delivered through the REAL fan-out
(tile_set._on_atlas_layers_received, never tile.on_response directly),
mirrored at single-window level. Two first-draft tests that passed with
the bug reverted were caught and strengthened before reporting; every
fix and hardening piece revert-verified independently (7+1 failing
tests without them).
Item 1: while zero tiles have arrived, the viewer draws a centered
screen-space 'DERIVING TERRAIN…' label (text_dim role, no new hue),
dropping the instant the first tile lands — a cold wait now reads as
loading, not broken. New has_any_tile_arrived() predicate (distinct
from has_pending_tiles(): both true mid-arrival, tested exactly there).
Item 2: legend re-fit root-caused empirically, two plausible fixes
disproven by trace before the real one: a manually-positioned Control's
size NEVER tracks a shrinking minimum in this parenting shape, and
RichTextLabel.fit_content reports degenerate minimums until laid out at
real width once — so reset must be DEFERRED and run after refill, not
inside clear(). ImplantPanel.reset_to_content_size() (call_deferred),
wired into both legend refresh()es. The load-bearing test compares
size.y to get_minimum_size().y — a size-to-size comparison passed
trivially with both numbers equally stuck (caught on first draft).
Item 3: make atlas now builds the RELEASE server and passes
SR_SERVER_BIN (a cold DEBUG server delivers zero tiles for >10s on a
new body — live-measured — vs 210ms warm; release serves cold in well
under a second). atlas_standalone._server_binary_path() honors the env
override per the SR_PORT two-tier precedent, debug path unchanged
when unset.
All fixes revert-verified; nine suites green collateral-checked;
gdlint clean.
Jeroen hit a black mosaic zooming into a body from a fresh make-atlas
spawn. Live diagnosis showed all six tiles held with colored textures
and no repaint; the original line-specific diagnosis (unpaired
queue_redraw in _on_tile_ready) turned out WRONG — the pairing already
existed (lead's truncated grep misread the function; Stig verified via
git log -p before acting). Rather than chase the exact dropped signal
edge, _process() now self-heals: both viewer and overlay redraw every
frame while the tile set has pending tiles (has_pending_tiles(), new) —
a strict superset that closes the black regardless of which edge drops,
pinned by a _draw()-counting real-subclass spy test.
Legend stacking root cause found by trace, not guess: ImplantApp.
_on_screen_changed() re-runs enter() unconditionally on repeat
same-screen pushes — each re-entry tore down and rebuilt all six tile
requests (the round-6 orphaning fingerprint via a new trigger) and
stacked another legend (~10 deep, full-height dark panel). Fixed both
ends: ImplantPanel.clear() frees immediately (same-frame re-entrant
refresh can never observe stale children — protects every implant app),
and RegionalScreen.enter() no-ops for the same body (different body
still re-enters fresh). The load-bearing regression asserts an ARRIVED
TILE'S DATA survives a repeat push — node identity would not catch the
teardown (the tile-set Node is a fixed field; only its internals reset).
Cold-start UX: pending tiles now draw the single-window path's
COLOR_BORDER_FADE wash instead of raw background — a deriving mosaic
reads as loading, not broken.
Suites green (zoom_ladder 48, viewer 74, tile_set 20, overlay 30,
overlays 46 + 2 new files), gdlint clean, revert-verified throughout.
All Hoshe/Araminta findings addressed (Tyre approved outright), none
retracted — plus a Dudley stop-and-flag discovery that improved on the
asked-for fix:
- Legend 100x lie (Araminta, blocking): subtitle computed via
spacing_for_rung() and refreshed at all three _held_granularity_v2
write sites. Region test asserts 204.800 km/cell; the District
direction needed a stale-header-aware helper — a District-only test
spuriously passes against the old literal by coincidence.
- Region coalescing coverage (Hoshe 1): both directions tested
(Region-vs-District separate slots; Region-vs-Region coalesces).
- Clamp boundary tests + dangling citations (Hoshe 2): writing the
requested halving-loop-fires test surfaced that the loop is PROVABLY
UNREACHABLE at current constants (per-axis clamp forecloses it —
brute-forced independently on both server and client sides). Ruling:
the loop stays as defensive code; the test became a property sweep
pinning both the wire-cap invariant and the loop's no-op status (a
future constant change breaks it loudly); doc comments on both sides
drop the load-bearing framing and state the truth; the old client
mirror test that claimed the loop fires (passing on the per-axis
clamp alone) is replaced the same way. Client citations now name the
real server tests verbatim.
- Governance (Tyre): D-226 amendment note — progressive cross-rung
refinement EXTENDS T-1124 §4 (not supersedes); legacy u32 field
scheduled for retirement (T-1159).
Server: 1818 lib tests green, clippy/fmt clean. Client: zoom_ladder
48/48, window_request 26/26, viewer 74/74; gdlint clean. Every fix
revert-verified.
The round-5 per-tick zoom comparison inside the reset guard was level-
triggered: at the canonical frame every continued zoom-out tick drifted
_view_zoom a hair below fit, read as 'not there', and re-fired
enter_orbital — tile set torn down/recreated per tick, in-flight
responses orphaned (black entry), server flooded (889 of 897 responses
in one drive phase). The drift concern now lives in _zoom_at as a ZOOM
FLOOR (once at the canonical frame, zoom-out clamps at fit); the guard
keys only on fields that never drift (_held_center/_held_granularity_v2/
_tile_mode) — genuinely edge-triggered. Three shared helpers keep the
guard and the floor on one definition of 'canonical frame'.
Anti-storm regression drives 100 real _zoom_at ticks past the first
reset and asserts tile-set child node identities never change; the
black-entry repro interleaves Lendel's six wire-accurate tile responses
with continued ticks and asserts all six held. Both revert-verified with
the exact live symptoms. Round-5 reset tests pass unchanged. Live-
verified twice: exactly 6 requests at entry and at reset, no storm.
The wrapped tile column (Lendel: -6400 canonicalized to 12739) drew at
its canonical column — off-canvas right — leaving the mosaic's left
third black. nearest_wrap_image() re-expresses a tile column as the
wrap-image closest to held_center for DRAWING only (requests/cache keys
stay canonical). The draw-position test asserts overlap FRACTION, not
bare intersects() — the buggy placement still clipped ~2px of viewport
edge at Lendel scale, so intersects() alone would false-pass.
The full-zoom-out reset restored center but not the fit zoom, and its
'already there' guard keyed on a lagging field so it could only ever
fire once. The guard now also matches tile mode and compares _view_zoom
against the freshly computed fit — restoring the FULL canonical
transform (center, offset, fit zoom) and re-firing as a continued
gesture keeps zooming out (Jeroen's hard condition, both live shapes).
+8 revert-verified tests incl. the continued-gesture reset repro; smoke
stub gained get_body_radius_km (crash confirmed real under a real
driver before fixing). Suites 286 green; gdlint clean.
Three live-found rendering bugs, all invisible to green unit suites:
- _draw_tile_mosaic placed tiles from absolute district (0,0); canvas-
local (0,0) is held_center - held_n/2 everywhere else, and tile-mode
held_n is the whole-body extent — the entire mosaic drew tens of
thousands of px off-canvas. New pure district_to_canvas_local() +
viewer accessors route every tile through the shared frame.
- _maybe_reselect_rung updated held_n across crossings without
recomputing _view_offset — the single-window composite landed off-
canvas the moment any crossing happened (why District/Quarter were
black too). New pure recompute_offset_for_held_n_change().
- _build_tile_texture created an unstored ImageTexture per _draw,
racing the RenderingServer's deferred upload — CPU pixels correct,
screen white. Per-tile-index texture cache, same reference-identity
discipline as the single-window _cached_texture.
Structural close of the twice-bitten 'nothing asserts pixels' gap:
test_atlas_window_overlay_draw_smoke.gd renders overlay output into a
SubViewport and asserts visible pixels for both modes — runs under a
real driver (invocation documented in DEVOPS.md, visual_capture
precedent; migration to the T-1157 harness noted on that ticket), skips
loud-but-green under the gate's headless run (verified green-with-skips
AND genuinely red with detection forced off). +7 geometry/crossing
tests, all revert-verified. Targeted suites 270 green; gdlint clean.
Bug A (zoom reselection dead): select_rung() redesigned from the two-gate
split to a unified per-rung coverage-ceiling walk (MAX_COVERAGE_M,
finest-first) — which also restores District as a reachable rung
(33-131km extent band; the two-gate model had made it structurally
unreachable). Plus the THIRD instance of the PR #191 C1 clamp-mirror bug
class: _maybe_reselect_rung fed the old rung's _held_n raw into the new
rung's request — the server clamped, the held value didn't, every
cross-rung response stale-dropped. Re-clamped at the boundary; the
leaving_tile_mode flag also fixes a stale request-granularity edge case.
Bug B (orbital = whole body): genuine tile-set model. compute_tile_grid
composes ceil-divided Region-tile centers (canonicalized column wrap,
row clamp, Dictionary-set dedup; Lendel = 3x2 = 6 tiles);
AtlasWindowTileSet owns one AtlasWindowRequest per tile (reusing all
tested request/cache/debounce machinery), fans responses out, per-tile
tile_ready — genuinely progressive; the overlay draws arrived tiles as a
mosaic under the one view transform. enter_orbital branches to tile mode
when >1 tile; single-window path unchanged below the coverage ceiling.
E/W wrap at tile seams live-confirmed (col -6400 wraps to 12739). Pole-
row dedup verified correct but flagged honestly as currently-unreachable
defensive logic (no real radius triggers row collision).
All fixes + fallout verified via revert/confirm-fails/restore. Targeted
suites 256/256; gdlint clean; new 20-test tile-set suite.
Second live-round blocker: the T-1150 legacy granularity comparison
stayed armed alongside the v2 check, and the server ALWAYS sends the
u32::MAX Region sentinel in the legacy slot — which can never equal the
client's pinned legacy value, so every Region response was stale-dropped
after the v2 check passed. _echoed_granularity_matches() now branches on
PRESENCE of granularity_v2: present -> v2 is the only comparison; absent
(old server) -> legacy fallback.
Mock-fleet audit while fixing: three tests had responses diverging from
the real wire — the oversized-orbital round-trip omitted the legacy
sentinel (passed for the wrong reason), the Region-accept test used
legacy=1, and the mismatch-drop test left v2 at a masking default that
would have inverted under the new rule. All wire-accurate now with a
named SERVER_LEGACY_GRANULARITY_REGION_SENTINEL const; +2 tests (legacy-
only old-server acceptance; v2-wins-regardless-of-legacy precedence
proof). All three verified to fail against the reverted fix. Full suite
3524/3524.
The live round against a real server caught what 3518 unit tests missed:
enter_orbital stored the UNCLAMPED n (Lendel: 19139 raw circumference
cols) into _held_n while AtlasWindowRequest clamped independently to
6400 before sending — every orbital response on a real-sized body was
dropped by the viewer's own w_n != _held_n staleness check. Permanently
blank ladder. The PR #191 C1 clamp-mirror lesson, one layer up: the
mirror must apply wherever held/expected state is recorded, not only at
the request layer.
_enter_at_rung now clamps via _clamp_window_n_mirror_v2 before storing
(the file's only _held_n write site — reselect/refloat read and forward).
Regression closes the coverage hole: the oversized-orbital RESPONSE
round-trip (enter_orbital on GJ380c -> deliver echo n=6400 Region ->
assert accepted). Both new tests verified to fail against the reverted
fix before trusting green. Full suite 3520/3520.
The atlas 'regional' screen now opens the LADDER at the canonical orbital
frame (Region granularity, whole body fitted and centered) and wheel zoom
descends continuously — cursor-anchored, unclamped across rungs, with
progressive refinement (held composite keeps drawing, finer rung swaps in
place on arrival; no blank frame, no mode flip). Full-zoom-out resets to
the canonical planetary frame per Jeroen's HARD condition
(is_fully_zoomed_out = extent >= body circumference, not a zoom-value
heuristic). The district_screen nav hop is deleted — D-013 restored:
descent is a zoom gesture, not a nav push. AtlasViewer's heightmap-texture
path is unreachable from nav (code intact; overlay surface deferred, see
report/tickets).
Rung selection: design doc §5's literal formula has NO legal District band
at any real viewport (visual-tolerance band and n=64 coverage ceiling
never overlap — pinned by executable boundary tests at 1600x900);
select_rung() splits it into a coverage ceiling (decides Region) then the
2x visual tolerance (District vs Quarter), documented at the function.
In practice the ladder steps Region -> Quarter directly.
Wire: window_granularity_v2 encoded (omitted at District for byte-compat),
granularity_v2 echoed value keyed + staleness-guarded end to end; Region
clamp mirror replicates the server's bounded halving loop (no closed
form). MIN/MAX_ZOOM widened to [0.0005, 64] — the old 0.5 floor would
have clamped a real body's canonical fit zoom, violating the reset
condition.
Real pre-existing bug fixed in atlas_window_overlay.gd: the draw path used
echoed n as both cell-grid dimension and district extent — only
coincidentally correct at District granularity; Quarter/Region would have
read wrong array offsets. cell_grid_side_for_window() now mirrors the
server's WindowGranularity::cell_grid_side.
Tests: +26 pure-function geometry tests, new 30-test zoom-ladder suite,
extensions across the window cache/request/overlay/delivery suites.
Full suite 3518 green; cold-parse clean.
All seven Hoshe/Tyre findings addressed, none retracted:
- n-clamp/echo/staleness triangle (Tyre C1): client _clamp_window_n_mirror
(bit-for-bit twin of the server clamp, canonicalize_district_center
precedent) applied before _n is stored/sent; server test pins the
quarter n=32 -> echo 16 contract.
- min_wl band quantization (Hoshe 1/Tyre C3): quantize_min_wl_m snaps to
MIN_WL_BANDS_M {0, 32768, 16384, 8192, 4096} before cache key and echo
(design doc §5's unbounded-key fix), reusing the one true
OCTAVE_WAVELENGTHS_M array; docstrings now state the server-quantizes/
client-sends-raw split; same-band cache-sharing test.
- coalescing granularity axis (Hoshe 2): two tests pin different-
granularity requests as separate in-flight slots and same-granularity
coalescing unchanged.
- orphaned fixture (Hoshe 3): test_protocol.gd consumer decodes
atlas_response_ready_with_window.msgpack through the real IPC path and
asserts the new fields.
- atlas_window_request coverage (Hoshe 4): new test file — stale-drop on
granularity mismatch, old-server-shape defaults accepted, clamp mirror
formula + wiring. First draft's quarter-via-request_now test would have
passed for the wrong reason (request_now resets granularity by design
until T-1153) — split into formula pin + reachable-path wiring proof.
- granularity type seam (Tyre C2): field + resolver docstrings state
finer-only integer multiples with resolve_window_granularity as the
single widening point; matching contract note added to the D-226
T-1143-rulings amendment.
cargo --lib 1807/1807; goldens bit-identical; gdlint clean.
T-1151: build_district_window_layer dispatches one Rayon task per row
(pure derive_window_cell via derive_at_metres), scattered row-major into
the flat arrays; a cfg(test) serial path backs the bit-identical
parallel-vs-serial golden.
T-1150: serde-default window_granularity (1=district, 4=quarter) +
window_min_wl_m on AtlasLayerRequest — additive, no sixth demux shape,
old frames decode unchanged (tested). Quarter mode = full
reclassification at 512m spacing over the same world rect ((4n)x(4n)
cells); WIRE_CAP_CELLS=4096 enforces n*granularity <= cap (quarter
clamps n to 16, the design doc's worked example). Granularity + min_wl
key ALL five touch points: DistrictWindowLayer echo, server FIFO-256
cache key (now a 5-tuple), per-connection coalescing key, client
request codec (omitted-at-default wire fields), client LRU key.
Mandatory aliasing regressions on both ends: identical (body, center, n)
at granularity 1 vs 4 produce distinct cache entries and correct
per-granularity payload shapes (server, 3-thread queue to avoid the
AnalyzeBody thread contention found while writing it) and distinct
client cache keys (gdUnit). Replay fixture regenerated — the layer
struct grew two echoed fields (231->254 bytes, content verified).
Client requests stay district-granularity by default — quarter requests
arrive with T-1153's rung selection.
Cover-fit: fit_window_view zooms from the viewport's LARGER dimension,
no margin factor (any factor under 1.0 leaves a long-axis gap — checked
numerically) — the composite fills edge to edge, overhanging the short
axis into pan-space; the refloat center-equality early-return already
prevents refetch churn at rest (proved, not just tested).
Input model (Jeroen: drag breaks click semantics with map objects):
LMB-drag pan REMOVED from the regional window; clicks are
object-reserved. Pan = held WASD/arrows polled in _process (delta- and
zoom-scaled, camera-pans-toward-key convention verified numerically)
plus edge-scroll within 24px of the viewport border; both suppressed
over UI and on OS focus loss; both set _user_adjusted; wheel zoom and
Esc unchanged. Reads RAW physical keycodes deliberately — independent
of the shared D-054 move_* InputMap actions bound to the same keys
(whose occlusion-leak is pre-existing and now ticketed as T-1146).
Pole wall + east-west wrap unchanged, re-driven through the new
inputs; drag tests replaced, not kept.
Smoothed composite (interim pending T-1143): per-cell colors bake into
an n x n Image/ImageTexture (exact existing colorizer incl. overlay +
ice tint) drawn once with LINEAR filtering — GPU bilinear reads as
terrain, the planetary heightmap's own treatment. Crisp per-cell path
preserved behind COMPOSITE_SMOOTH for T-1143 A/B. Rebuild only on
reference-identity change of window/toggle (is_same — verified true
reference equality; value-equal distinct dicts DO rebuild).
Governance: T-1145 amendment paragraph on D-226 T-1124 SS5 (all three
supersessions); pql decisions validate ok.
Suites: window_viewer 74/74, window_geometry 32/32, window_overlay
(new) 16/16; gdlint clean on all six files.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Bounds gate: AtlasDescendGeometry.is_on_texture() — ONE helper feeding
both the reticle guard and the click fall-through (the T-1140 lesson:
the visible affordance always matches the click); half-open
[0,tex_w)x[0,tex_h) boundary pinned at the exact edge. Letterbox
clicks no longer show a reticle or descend.
Fit-and-center: pure fit_window_view() (new atlas_window_geometry.gd)
wired into enter(), the FIRST window arrival, and NOTIFICATION_RESIZED
— gated by a _user_adjusted flag so the fit never fights manual
zoom/pan (flag clears only on a fresh enter). Found-own-bug: RESIZED
can fire mid-_ready() before _canvas exists — null-guarded like the
sibling panels.
Pole wall (Jeroen's ruling): clamp_pan_offset_to_pole_wall() clamps
the WINDOW EDGE, not the center, in screen space from the fitted
transform — Y only; wired into the drag handler and every fit (a
fresh fit can itself need the wall on a tiny body — the
window-taller-than-planet case is handled and tested). Three numeric
hand-traces preceded the code; a first-draft test using GJ380c's huge
radius silently never exercised the clamp — replaced with a synthetic
small radius.
East-west wrap (Jeroen's ruling): canonicalize_district_center() —
posmod column wrap (verified against a live Godot process to match
Rust rem_euclid bit-for-bit), clamped row; district_extent() shares
the exact formula (incl. .max(1)) with the server's
normalize_window_center so echoes and cache keys agree on canonical
form. Canonicalization applies only to the FINAL refetch center — the
edge-crossing decision stays in absolute district space (first-pass
math error caught by hand-trace). Seam-adjacent cache-key sharing
tested. Pan offset itself has no x wall — circumnavigation is
seamless.
Header: body proper_name/body_id ahead of the coordinates (the cheap
half of T-1141, noted in code). Drag-pan verified through the REAL
DistrictScreen-to-viewer chain and pinned by test (no fix needed).
140 tests across three suites, 0 failures; 94 sibling tests no
ripple; gdlint clean (atlas_viewer.gd at the 1000-line cap a second
round — structural extraction flagged for maintenance).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Hoshe's blocking find: district_pos_at divided the latitude axis by
tex_h where the server forward map uses ta.h.saturating_sub(1)
(district_profile.rs:1436; ground-truth inverse aliveness_probe.rs:511)
— every off-equator click descended into the wrong district, up to
~20km drift, live-repro'd at 10.2km. Fixed to /(tex_h - 1.0) with the
reference implementation's degenerate-texture guard; verified against
the repro (row 50: old formula recovered 41, fixed recovers 50 exact).
Columns were already correct (longitude wraps — the asymmetry the
docstring documented but the code didn't implement).
Regression tests transcribe the server forward formula (source cited)
and round-trip three off-equator rows — one per hemisphere plus
near-pole — asserting EXACT recovery; a dedicated drift guard is
framed as Tyre's C2 (a client-side twin of a server mapping owns its
own round-trip proof). Root-caused the coverage hole: the old
equator test sampled tex_h*0.5 believing it was the equator pixel —
the true equator is (tex_h-1)*0.5, so the test never verified what it
claimed; corrected with the why documented.
Araminta's find: the descend reticle drew unconditionally on hover —
over a city the marker flared white (city-click signal) while the
reticle+extent label promised descent, though the click correctly
resolved to city data. Extracted _should_draw_descend_reticle() with
the missing _hovered_city.is_empty() clause; three state tests pin
shows-over-map / hides-over-city / hides-when-idle.
test_atlas_descend_entry 32/32; adjacent clusters no ripple; gdlint
clean (atlas_viewer.gd at exactly the 1000-line cap).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Entry per Jeroen's 2026-07-21 revision: planetary heightmap is now
FIXED — all drag-pan/wheel-zoom input removed (set_view/get_view_*
capture API survives for the golden harness); hover shows a
not-to-scale bracket reticle with the real extent labeled (a true
n=32 rectangle is sub-pixel on the planetary canvas — the honest
representation given the morph transition is deferred), and a click
that misses every city marker descends (city-click wins — one
gesture, two contextual reads, no modifier). Descent pushes a new
'district' nav screen centered on the click point's DistrictPos via
atlas_descend_geometry.district_pos_at (the pixel-to-district inverse
of the server mapping, verified against scale.rs).
Regional mode: atlas_window_viewer draws the composite (morphology x
elev_q lightness base; temp/moisture/veg toggles — temp reuses the
region-ramp colorizer exactly; Marine=6 transparent; glaciation
always-on tint matching apply_ice_tint's REAL gate, None|Light no-op,
over the amendment's looser prose — documented); pan-on-held-composite
with edge-crossing refetch + border-fade during the queue-based
derive wait; zoom never refetches. atlas_window_cache: LRU keyed
(body_id, center, n), touch-on-read, evict-only, no freshness (D-227).
atlas_window_request mirrors the generation-proxy pending-retry shape
for None-until-derived. Codec: window params omitted from the wire
when absent — byte-identical for every existing caller.
Live-verified against the T-1137 server in-worktree: real round-trip
on a GJ380c coastal district (6 fields x 1024 cells), echo staleness
guard, genuine ~1.4s background-derive wait, pan-edge refetch to an
adjacent window, cache-hit on re-descent with zero network. Full
client suite 3194/3194; gdlint clean on all 18 files.
Open follow-ups flagged in-code: header location label always falls
back to coordinates (nearest-settlement needs a join the district
window does not carry); atlas_standalone.gd's 'atlas_app.gd is never
modified' doc line is now imprecise.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
